Closed LiuXuan66 closed 1 year ago
imagenrom:5.27702ms camera backbone:29.77021ms BEV vision transform :1.94419ms BEV fuse head: 17.78800ms BEV postprocess: 5.88835ms Total: 55.391ms
I try again on another new Orin mechine. imagenrom:4.20195ms camera backbone:12.36874ms BEV vision transform :1.30166ms BEV fuse head: 69.65811ms BEV postprocess: 4.74912ms Total: 88.078ms
I just test on 2080Ti , no orin,maybe you can use TensorRT8.6 ,and rebuild int8 model, version and hardware affect the optimization.Such a large amount of time should not be spent on lightweight backbones and fuse necks
Thanks a lot. I will try to update the version of tensorrt.
the time consumption on orin platform is as below
==================FastBEV===================
[⏰ [NoSt] ImageNrom]: 3.01830 ms
[⏰ Camera Backbone]: 8.91315 ms
[⏰ BEV vision transform]: 0.71488 ms
[⏰ BEV fuse head]: 66.66816 ms
[⏰ BEV postprocess]: 4.80109 ms
Total: 81.097 ms
=============================================
==================FastBEV===================
[⏰ [NoSt] ImageNrom]: 3.17738 ms
[⏰ Camera Backbone]: 8.15795 ms
[⏰ BEV vision transform]: 0.71149 ms
[⏰ BEV fuse head]: 66.38195 ms
[⏰ BEV postprocess]: 4.75507 ms
Total: 80.006 ms
=============================================
==================FastBEV===================
[⏰ [NoSt] ImageNrom]: 3.16928 ms
[⏰ Camera Backbone]: 8.14246 ms
[⏰ BEV vision transform]: 1.20810 ms
[⏰ BEV fuse head]: 66.38873 ms
[⏰ BEV postprocess]: 4.76374 ms
Total: 80.503 ms
=============================================
==================FastBEV===================
[⏰ [NoSt] ImageNrom]: 3.15174 ms
[⏰ Camera Backbone]: 8.14074 ms
[⏰ BEV vision transform]: 1.21091 ms
[⏰ BEV fuse head]: 66.38090 ms
[⏰ BEV postprocess]: 4.75018 ms
Total: 80.483 ms
=============================================
==================FastBEV===================
[⏰ [NoSt] ImageNrom]: 3.15520 ms
[⏰ Camera Backbone]: 8.14307 ms
[⏰ BEV vision transform]: 1.21088 ms
[⏰ BEV fuse head]: 66.37798 ms
[⏰ BEV postprocess]: 4.73478 ms
Total: 80.467 ms
=============================================
==================FastBEV===================
[⏰ [NoSt] ImageNrom]: 3.12976 ms
[⏰ Camera Backbone]: 8.14323 ms
[⏰ BEV vision transform]: 1.20963 ms
[⏰ BEV fuse head]: 66.39207 ms
[⏰ BEV postprocess]: 4.73328 ms
Total: 80.478 ms
=============================================
is there something wrong?
我在orin上耗时:
==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 4.80746 ms [⏰ Camera Backbone]: 16.43366 ms [⏰ BEV vision transform]: 1.10966 ms [⏰ BEV fuse head]: 11.59578 ms [⏰ BEV postprocess]: 5.07693 ms Total: 34.216 ms
`
int8的。
the time consumption on orin platform is as below
==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 3.01830 ms [⏰ Camera Backbone]: 8.91315 ms [⏰ BEV vision transform]: 0.71488 ms [⏰ BEV fuse head]: 66.66816 ms [⏰ BEV postprocess]: 4.80109 ms Total: 81.097 ms ============================================= ==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 3.17738 ms [⏰ Camera Backbone]: 8.15795 ms [⏰ BEV vision transform]: 0.71149 ms [⏰ BEV fuse head]: 66.38195 ms [⏰ BEV postprocess]: 4.75507 ms Total: 80.006 ms ============================================= ==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 3.16928 ms [⏰ Camera Backbone]: 8.14246 ms [⏰ BEV vision transform]: 1.20810 ms [⏰ BEV fuse head]: 66.38873 ms [⏰ BEV postprocess]: 4.76374 ms Total: 80.503 ms ============================================= ==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 3.15174 ms [⏰ Camera Backbone]: 8.14074 ms [⏰ BEV vision transform]: 1.21091 ms [⏰ BEV fuse head]: 66.38090 ms [⏰ BEV postprocess]: 4.75018 ms Total: 80.483 ms ============================================= ==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 3.15520 ms [⏰ Camera Backbone]: 8.14307 ms [⏰ BEV vision transform]: 1.21088 ms [⏰ BEV fuse head]: 66.37798 ms [⏰ BEV postprocess]: 4.73478 ms Total: 80.467 ms ============================================= ==================FastBEV=================== [⏰ [NoSt] ImageNrom]: 3.12976 ms [⏰ Camera Backbone]: 8.14323 ms [⏰ BEV vision transform]: 1.20963 ms [⏰ BEV fuse head]: 66.39207 ms [⏰ BEV postprocess]: 4.73328 ms Total: 80.478 ms =============================================
is there something wrong?
HI Have you solved this problem? Thanks. @ASONG0506 @QiTianDaShengDaShi @LiuXuan66
Hi, I just use tensorrt 8.5.2 to export the provided onnx model to trt on Orin . However, the inference time of int8 resnet18 model reaches 60ms per frame, fp16 reaches 100ms . I try to find reasons but failed. (eg. set 'release' not 'debug',update the tensorrt)